TEST(DSPSingle, TestConvolve)
{
float out[121];
float in1[5] =
{
0.0, 0.5, 1.0, 0.5, 0.0
};
float in2[6] =
{
0.5, 0.5, 0.5, 1.0, 1.0, 1.0
};
// >> conv([0 0.5 1 0.5 0], [0.5 0.5 0.5 1.0 1.0 1.0])
float res[10] =
{
0.0, 0.25, 0.75, 1.0, 1.25, 1.75, 2.0, 1.5, 0.5, 0.0
};
Convolve(in1, 5, in2, 6, out);
for (unsigned i = 0; i < 10; ++i)
{
ASSERT_FLOAT_EQ(res[i], out[i]);
}
}
void ImageProximityFFT::SqrDistance(Image& Source, Image& Template, Image& Dest)
{
CheckFloat(Dest);
CheckSameNbChannels(Source, Template);
CheckSameNbChannels(Source, Dest);
CheckSameSize(Source, Dest);
// Verify image size
if (Template.Width() > Source.Width() || Template.Height() > Source.Height())
throw cl::Error(CL_IMAGE_FORMAT_NOT_SUPPORTED, "The template image must be smaller than source image.");
// Verify image types
if(!SameType(Source, Template))
throw cl::Error(CL_IMAGE_FORMAT_MISMATCH, "The source image and the template image must be same type.");
PrepareFor(Source, Template);
m_integral.SqrIntegral(Source, *m_image_sqsums);
double templ_sqsum[4] = {0};
m_statistics.SumSqr(Template, templ_sqsum);
Convolve(Source, Template, Dest); // Computes the cross correlation using FFT
MatchSquareDiff(Template.Width(), Template.Height(), *m_image_sqsums, templ_sqsum, Dest);
}
void ConvolveSeparable(CImageOf<T> src, CImageOf<T>& dst,
CFloatImage x_kernel, CFloatImage y_kernel,
float scale, float offset,
int decimate, int interpolate)
{
// Allocate the result, if necessary
CShape dShape = src.Shape();
if (decimate > 1)
{
dShape.width = (dShape.width + decimate-1) / decimate;
dShape.height = (dShape.height + decimate-1) / decimate;
}
dst.ReAllocate(dShape, false);
// Allocate the intermediate images
CImageOf<T> tmpImg1(src.Shape());
CImageOf<T> tmpImg2(src.Shape());
// Create a proper vertical convolution kernel
CFloatImage v_kernel(1, y_kernel.Shape().width, 1);
for (int k = 0; k < y_kernel.Shape().width; k++)
v_kernel.Pixel(0, k, 0) = y_kernel.Pixel(k, 0, 0);
v_kernel.origin[1] = y_kernel.origin[0];
// Perform the two convolutions
Convolve(src, tmpImg1, x_kernel, 1.0f, 0.0f);
Convolve(tmpImg1, tmpImg2, v_kernel, scale, offset);
// Downsample or copy
for (int y = 0; y < dShape.height; y++)
{
T* sPtr = &tmpImg2.Pixel(0, y * decimate, 0);
T* dPtr = &dst.Pixel(0, y, 0);
int nB = dShape.nBands;
for (int x = 0; x < dShape.width; x++)
{
for (int b = 0; b < nB; b++)
dPtr[b] = sPtr[b];
sPtr += decimate * nB;
dPtr += nB;
}
}
interpolate++; // to get rid of "unused parameter" warning
}
void GetHarrisComponents(CFloatImage &srcImage, CFloatImage &A, CFloatImage &B, CFloatImage &C, CFloatImage *partialX, CFloatImage *partialY)
{
int w = srcImage.Shape().width;
int h = srcImage.Shape().height;
CFloatImage *partialXPtr;
CFloatImage *partialYPtr;
if (partialX != nullptr && partialY != nullptr)
{
partialXPtr = partialX;
partialYPtr = partialY;
}
else
{
partialXPtr = new CFloatImage(srcImage.Shape());
partialYPtr = new CFloatImage(srcImage.Shape());
}
CFloatImage partialXX(srcImage.Shape());
CFloatImage partialYY(srcImage.Shape());
CFloatImage partialXY(srcImage.Shape());
CFloatImage gaussianImage = GetImageFromMatrix((float *)gaussian5x5Float, 5, 5);
Convolve(srcImage, *partialXPtr, ConvolveKernel_SobelX);
Convolve(srcImage, *partialYPtr, ConvolveKernel_SobelY);
for (int y = 0; y < h; y++) {
for (int x = 0; x < w; x++) {
float *xxPixel = &partialXX.Pixel(x, y, 0);
float *yyPixel = &partialYY.Pixel(x, y, 0);
float *xyPixel = &partialXY.Pixel(x, y, 0);
// The 1/8 factor is to do the scaling inherent in sobel filtering
*xxPixel = pow((double)(1./8. *8. * partialXPtr->Pixel(x, y, 0)), 2.);
*yyPixel = pow((double)(1./8. *8. * partialYPtr->Pixel(x, y, 0)), 2.);
*xyPixel = pow(1./8. *8., 2.) * partialXPtr->Pixel(x, y, 0) * partialYPtr->Pixel(x, y, 0);
}
}
Convolve(partialXX, A, gaussianImage);
Convolve(partialXY, B, gaussianImage);
Convolve(partialYY, C, gaussianImage);
}
void ConvolveSeparable(CImageOf<T> src, CImageOf<T>& dst,
CFloatImage x_kernel, CFloatImage y_kernel,
int subsample)
{
// Allocate the result, if necessary
CShape dShape = src.Shape();
if (subsample > 1)
{
dShape.width = (dShape.width + subsample-1) / subsample;
dShape.height = (dShape.height + subsample-1) / subsample;
}
dst.ReAllocate(dShape, false);
// Allocate the intermediate images
CImageOf<T> tmpImg1(src.Shape());
CImageOf<T> tmpImg2(src.Shape());
// Create a proper vertical convolution kernel
CFloatImage v_kernel(1, y_kernel.Shape().width, 1);
for (int k = 0; k < y_kernel.Shape().width; k++)
v_kernel.Pixel(0, k, 0) = y_kernel.Pixel(k, 0, 0);
v_kernel.origin[1] = y_kernel.origin[0];
// Perform the two convolutions
Convolve(src, tmpImg1, x_kernel);
Convolve(tmpImg1, tmpImg2, v_kernel);
// Downsample or copy
for (int y = 0; y < dShape.height; y++)
{
T* sPtr = &tmpImg2.Pixel(0, y * subsample, 0);
T* dPtr = &dst.Pixel(0, y, 0);
int nB = dShape.nBands;
for (int x = 0; x < dShape.width; x++)
{
for (int b = 0; b < nB; b++)
dPtr[b] = sPtr[b];
sPtr += subsample * nB;
dPtr += nB;
}
}
}
bool LarsonSekaninaInstance::ExecuteOn( View& view )
{
AutoViewLock lock( view );
ImageVariant image = view.Image();
if ( image.IsComplexSample() )
return false;
StandardStatus status;
image.SetStatusCallback( &status );
Console().EnableAbort();
ImageVariant sharpImg;
sharpImg.CreateFloatImage( (image.BitsPerSample() > 32) ? image.BitsPerSample() : 32 );
sharpImg.AllocateImage( image->Width(), image->Height(), 1, ColorSpace::Gray );
if ( useLuminance && image->IsColor() )
{
ImageVariant L;
image.GetLightness( L );
Convolve( L, sharpImg, interpolation, radiusDiff, angleDiff, center, 0 );
ApplyFilter( L, sharpImg, amount, threshold, deringing, rangeLow, rangeHigh, false, 0, highPass );
image.SetLightness( L );
}
else
{
for ( int c = 0, n = image->NumberOfNominalChannels(); c < n; ++c )
{
image->SelectChannel( c );
if ( n > 1 )
Console().WriteLn( "<end><cbr>Processing channel #" + String( c ) );
Convolve( image, sharpImg, interpolation, radiusDiff, angleDiff, center, c );
ApplyFilter( image, sharpImg, amount, threshold, deringing, rangeLow, rangeHigh, disableExtension, c, highPass );
}
}
return true;
}
static void Apply( GenericImage<P>& image, const FFTConvolution& F )
{
Rect r = image.SelectedRectangle();
if ( F.m_h.IsNull() )
if ( !F.m_filter.IsNull() )
F.m_h = Initialize( *F.m_filter, r.Width(), r.Height(), F.IsParallelProcessingEnabled(), F.MaxProcessors() );
else
F.m_h = Initialize( F.m_image, r.Width(), r.Height(), F.IsParallelProcessingEnabled(), F.MaxProcessors() );
Convolve( image, *F.m_h, F.IsParallelProcessingEnabled(), F.MaxProcessors() );
}
LPyramid::LPyramid(const float *image, unsigned int width, unsigned int height) :
Width(width),
Height(height)
{
// Make the Laplacian pyramid by successively
// copying the earlier levels and blurring them
for (unsigned int i = 0; i < MAX_PYR_LEVELS; i++) {
if (i == 0) {
Levels[i] = Copy(image);
} else {
Levels[i] = new float[Width * Height];
Convolve(Levels[i], Levels[i - 1]);
}
}
}
void Image::Sharpen()
{
int* filt = new int[9]; //sharpening filter
filt[0] = -1;
filt[1] = -2;
filt[2] = -1;
filt[3] = -2;
filt[4] = 19;
filt[5] = -2;
filt[6] = -1;
filt[7] = -2;
filt[8] = -1;
int norm = 7;
int n = 3;
Convolve(filt, n, norm, false);
}
void ImageProximityFFT::CrossCorr(Image& Source, Image& Template, Image& Dest)
{
CheckFloat(Dest);
CheckSameNbChannels(Source, Template);
CheckSameNbChannels(Source, Dest);
CheckSameSize(Source, Dest);
// Verify image size
if (Template.Width() > Source.Width() || Template.Height() > Source.Height())
throw cl::Error(CL_IMAGE_FORMAT_NOT_SUPPORTED, "The template image must be smaller than source image.");
// Verify image types
if(!SameType(Source, Template))
throw cl::Error(CL_IMAGE_FORMAT_MISMATCH, "The source image and the template image must be same type.");
PrepareFor(Source, Template);
Convolve(Source, Template, Dest); // Computes the cross correlation using FFT
}
void Image::Blur(int n)
{
/* Your Work Here (Section 3.4.1) */
double sig = floor(n / (double) 2) / 2;
int mp = int(n / 2);
double* filtU = new double[n]; //f(u)
for (int i = -mp; i <= mp; i++) {
filtU[i + mp] = (1 / (sqrt(2 * M_PI) * sig)) * exp(-(i*i) / (2 * sig*sig));
}
double norm = 0.0;
int* filt = new int[n*n]; //blur filter
for (int i = 0; i < n; i++) {
for (int j = 0; j < n; j++) {
int temp = (int) ((filtU[i] * filtU[j]) / (filtU[0] * filtU[0]));
filt[i + j*n] = temp;
norm += temp;
}
}
Convolve(filt, n, (int) norm, false);
}
void subsample(Feature* f, int imgSize, CFloatImage gaussianImage)
{
vector<double, std::allocator<double>>::iterator it;
CFloatImage img = featureToImage(*f, imgSize, imgSize);
CFloatImage blurredImg(img.Shape());
Convolve(img, blurredImg, gaussianImage);
featuresFromImage(f,blurredImg,imgSize,imgSize);
int count = 0;
for(int y=0; y<imgSize; y++)
{
for(int x=0; x<imgSize; x++)
{
if(x%2 == 0 || y%2 == 0)
{
f->data.erase(f->data.begin() + count);
}
else
{
count++;
}
}
}
}
/*************************************************************************
*
* Function: cl_ltp
* Purpose: closed-loop fractional pitch search
*
**************************************************************************
*/
void cl_ltp (
clLtpState *clSt, /* i/o : State struct */
tonStabState *tonSt, /* i/o : State struct */
enum Mode mode, /* i : coder mode */
Word16 frameOffset, /* i : Offset to subframe */
Word16 T_op[], /* i : Open loop pitch lags */
Word16 *h1, /* i : Impulse response vector Q12 */
Word16 *exc, /* i/o : Excitation vector Q0 */
Word16 res2[], /* i/o : Long term prediction residual Q0 */
Word16 xn[], /* i : Target vector for pitch search Q0 */
Word16 lsp_flag, /* i : LSP resonance flag */
Word16 xn2[], /* o : Target vector for codebook search Q0 */
Word16 y1[], /* o : Filtered adaptive excitation Q0 */
Word16 *T0, /* o : Pitch delay (integer part) */
Word16 *T0_frac, /* o : Pitch delay (fractional part) */
Word16 *gain_pit, /* o : Pitch gain Q14 */
Word16 g_coeff[], /* o : Correlations between xn, y1, & y2 */
Word16 **anap, /* o : Analysis parameters */
Word16 *gp_limit /* o : pitch gain limit */
)
{
Word16 i;
Word16 index;
Word32 L_temp; /* temporarily variable */
Word16 resu3; /* flag for upsample resolution */
Word16 gpc_flag;
/*----------------------------------------------------------------------*
* Closed-loop fractional pitch search *
*----------------------------------------------------------------------*/
*T0 = Pitch_fr(&clSt->pitchSt,
mode, T_op, exc, xn, h1,
L_SUBFR, frameOffset,
T0_frac, &resu3, &index);
*(*anap)++ = index;
/*-----------------------------------------------------------------*
* - find unity gain pitch excitation (adapitve codebook entry) *
* with fractional interpolation. *
* - find filtered pitch exc. y1[]=exc[] convolve with h1[]) *
* - compute pitch gain and limit between 0 and 1.2 *
* - update target vector for codebook search *
* - find LTP residual. *
*-----------------------------------------------------------------*/
Pred_lt_3or6(exc, *T0, *T0_frac, L_SUBFR, resu3);
Convolve(exc, h1, y1, L_SUBFR);
/* gain_pit is Q14 for all modes */
*gain_pit = G_pitch(mode, xn, y1, g_coeff, L_SUBFR);
/* check if the pitch gain should be limit due to resonance in LPC filter */
gpc_flag = 0;
*gp_limit = MAX_16;
if (lsp_flag != 0 && *gain_pit > GP_CLIP)
{
gpc_flag = check_gp_clipping(tonSt, *gain_pit);
}
/* special for the MR475, MR515 mode; limit the gain to 0.85 to */
/* cope with bit errors in the decoder in a better way. */
if (mode == MR475 || mode == MR515) {
if (*gain_pit > 13926) {
*gain_pit = 13926; /* 0.85 in Q14 */
}
if (gpc_flag != 0) {
*gp_limit = GP_CLIP;
}
}
else
{
if (gpc_flag != 0)
{
*gp_limit = GP_CLIP;
*gain_pit = GP_CLIP;
}
/* For MR122, gain_pit is quantized here and not in gainQuant */
if ( mode == MR122 )
{
*(*anap)++ = q_gain_pitch(MR122, *gp_limit, gain_pit,
NULL, NULL);
}
}
/* update target vector und evaluate LTP residual */
for (i = 0; i < L_SUBFR; i++) {
L_temp = ((Word32)y1[i] * *gain_pit) >> 14;
xn2[i] = xn[i] - (Word16)L_temp;
L_temp = ((Word32)exc[i] * *gain_pit) >> 14;
res2[i] -= (Word16)L_temp;
}
}
void LinearFilter::Invoke()
{
Convolve();
//Display(1);
}
/***************************************************************************
* FUNCTION: cod_amr
*
* PURPOSE: Main encoder routine.
*
* DESCRIPTION: This function is called every 20 ms speech frame,
* operating on the newly read 160 speech samples. It performs the
* principle encoding functions to produce the set of encoded parameters
* which include the LSP, adaptive codebook, and fixed codebook
* quantization indices (addresses and gains).
*
* INPUTS:
* No input argument are passed to this function. However, before
* calling this function, 160 new speech data should be copied to the
* vector new_speech[]. This is a global pointer which is declared in
* this file (it points to the end of speech buffer minus 160).
*
* OUTPUTS:
*
* ana[]: vector of analysis parameters.
* synth[]: Local synthesis speech (for debugging purposes)
*
***************************************************************************/
int cod_amr(
cod_amrState *st, /* i/o : State struct */
enum Mode mode, /* i : AMR mode */
Word16 new_speech[], /* i : speech input (L_FRAME) */
Word16 ana[], /* o : Analysis parameters */
enum Mode *usedMode, /* o : used mode */
Word16 synth[] /* o : Local synthesis */
)
{
/* LPC coefficients */
Word16 A_t[(MP1) * 4]; /* A(z) unquantized for the 4 subframes */
Word16 Aq_t[(MP1) * 4]; /* A(z) quantized for the 4 subframes */
Word16 *A, *Aq; /* Pointer on A_t and Aq_t */
Word16 lsp_new[M];
/* Other vectors */
Word16 xn[L_SUBFR]; /* Target vector for pitch search */
Word16 xn2[L_SUBFR]; /* Target vector for codebook search */
Word16 code[L_SUBFR]; /* Fixed codebook excitation */
Word16 y1[L_SUBFR]; /* Filtered adaptive excitation */
Word16 y2[L_SUBFR]; /* Filtered fixed codebook excitation */
Word16 gCoeff[6]; /* Correlations between xn, y1, & y2: */
Word16 res[L_SUBFR]; /* Short term (LPC) prediction residual */
Word16 res2[L_SUBFR]; /* Long term (LTP) prediction residual */
/* Vector and scalars needed for the MR475 */
Word16 xn_sf0[L_SUBFR]; /* Target vector for pitch search */
Word16 y2_sf0[L_SUBFR]; /* Filtered codebook innovation */
Word16 code_sf0[L_SUBFR]; /* Fixed codebook excitation */
Word16 h1_sf0[L_SUBFR]; /* The impulse response of sf0 */
Word16 mem_syn_save[M]; /* Filter memory */
Word16 mem_w0_save[M]; /* Filter memory */
Word16 mem_err_save[M]; /* Filter memory */
Word16 sharp_save; /* Sharpening */
Word16 evenSubfr; /* Even subframe indicator */
Word16 T0_sf0 = 0; /* Integer pitch lag of sf0 */
Word16 T0_frac_sf0 = 0; /* Fractional pitch lag of sf0 */
Word16 i_subfr_sf0 = 0; /* Position in exc[] for sf0 */
Word16 gain_pit_sf0; /* Quantized pitch gain for sf0 */
Word16 gain_code_sf0; /* Quantized codebook gain for sf0 */
/* Scalars */
Word16 i_subfr, subfrNr;
Word16 T_op[L_FRAME/L_FRAME_BY2];
Word16 T0, T0_frac;
Word16 gain_pit, gain_code;
/* Flags */
Word16 lsp_flag = 0; /* indicates resonance in LPC filter */
Word16 gp_limit; /* pitch gain limit value */
Word16 vad_flag; /* VAD decision flag */
Word16 compute_sid_flag; /* SID analysis flag */
Copy(new_speech, st->new_speech, L_FRAME);
*usedMode = mode; move16 ();
/* DTX processing */
if (st->dtx)
{ /* no test() call since this if is only in simulation env */
/* Find VAD decision */
#ifdef VAD2
vad_flag = vad2 (st->new_speech, st->vadSt);
vad_flag = vad2 (st->new_speech+80, st->vadSt) || vad_flag; logic16();
#else
vad_flag = vad1(st->vadSt, st->new_speech);
#endif
fwc (); /* function worst case */
/* NB! usedMode may change here */
compute_sid_flag = tx_dtx_handler(st->dtx_encSt,
vad_flag,
usedMode);
}
else
{
compute_sid_flag = 0; move16 ();
}
/*------------------------------------------------------------------------*
* - Perform LPC analysis: *
* * autocorrelation + lag windowing *
* * Levinson-durbin algorithm to find a[] *
* * convert a[] to lsp[] *
* * quantize and code the LSPs *
* * find the interpolated LSPs and convert to a[] for all *
* subframes (both quantized and unquantized) *
*------------------------------------------------------------------------*/
/* LP analysis */
lpc(st->lpcSt, mode, st->p_window, st->p_window_12k2, A_t);
fwc (); /* function worst case */
/* From A(z) to lsp. LSP quantization and interpolation */
lsp(st->lspSt, mode, *usedMode, A_t, Aq_t, lsp_new, &ana);
fwc (); /* function worst case */
/* Buffer lsp's and energy */
dtx_buffer(st->dtx_encSt,
lsp_new,
st->new_speech);
/* Check if in DTX mode */
test();
if (sub(*usedMode, MRDTX) == 0)
{
dtx_enc(st->dtx_encSt,
compute_sid_flag,
st->lspSt->qSt,
st->gainQuantSt->gc_predSt,
&ana);
Set_zero(st->old_exc, PIT_MAX + L_INTERPOL);
Set_zero(st->mem_w0, M);
Set_zero(st->mem_err, M);
Set_zero(st->zero, L_SUBFR);
Set_zero(st->hvec, L_SUBFR); /* set to zero "h1[-L_SUBFR..-1]" */
/* Reset lsp states */
lsp_reset(st->lspSt);
Copy(lsp_new, st->lspSt->lsp_old, M);
Copy(lsp_new, st->lspSt->lsp_old_q, M);
/* Reset clLtp states */
cl_ltp_reset(st->clLtpSt);
st->sharp = SHARPMIN; move16 ();
}
else
{
/* check resonance in the filter */
lsp_flag = check_lsp(st->tonStabSt, st->lspSt->lsp_old); move16 ();
}
/*----------------------------------------------------------------------*
* - Find the weighted input speech w_sp[] for the whole speech frame *
* - Find the open-loop pitch delay for first 2 subframes *
* - Set the range for searching closed-loop pitch in 1st subframe *
* - Find the open-loop pitch delay for last 2 subframes *
*----------------------------------------------------------------------*/
#ifdef VAD2
if (st->dtx)
{ /* no test() call since this if is only in simulation env */
st->vadSt->L_Rmax = 0; move32 ();
st->vadSt->L_R0 = 0; move32 ();
}
#endif
for(subfrNr = 0, i_subfr = 0;
subfrNr < L_FRAME/L_FRAME_BY2;
subfrNr++, i_subfr += L_FRAME_BY2)
{
/* Pre-processing on 80 samples */
pre_big(mode, gamma1, gamma1_12k2, gamma2, A_t, i_subfr, st->speech,
st->mem_w, st->wsp);
test (); test ();
if ((sub(mode, MR475) != 0) && (sub(mode, MR515) != 0))
{
/* Find open loop pitch lag for two subframes */
ol_ltp(st->pitchOLWghtSt, st->vadSt, mode, &st->wsp[i_subfr],
&T_op[subfrNr], st->old_lags, st->ol_gain_flg, subfrNr,
st->dtx);
}
}
fwc (); /* function worst case */
test (); test();
if ((sub(mode, MR475) == 0) || (sub(mode, MR515) == 0))
{
/* Find open loop pitch lag for ONE FRAME ONLY */
/* search on 160 samples */
ol_ltp(st->pitchOLWghtSt, st->vadSt, mode, &st->wsp[0], &T_op[0],
st->old_lags, st->ol_gain_flg, 1, st->dtx);
T_op[1] = T_op[0]; move16 ();
}
fwc (); /* function worst case */
#ifdef VAD2
if (st->dtx)
{ /* no test() call since this if is only in simulation env */
LTP_flag_update(st->vadSt, mode);
}
#endif
#ifndef VAD2
/* run VAD pitch detection */
if (st->dtx)
{ /* no test() call since this if is only in simulation env */
vad_pitch_detection(st->vadSt, T_op);
}
#endif
fwc (); /* function worst case */
if (sub(*usedMode, MRDTX) == 0)
{
goto the_end;
}
/*------------------------------------------------------------------------*
* Loop for every subframe in the analysis frame *
*------------------------------------------------------------------------*
* To find the pitch and innovation parameters. The subframe size is *
* L_SUBFR and the loop is repeated L_FRAME/L_SUBFR times. *
* - find the weighted LPC coefficients *
* - find the LPC residual signal res[] *
* - compute the target signal for pitch search *
* - compute impulse response of weighted synthesis filter (h1[]) *
* - find the closed-loop pitch parameters *
* - encode the pitch dealy *
* - update the impulse response h1[] by including fixed-gain pitch *
* - find target vector for codebook search *
* - codebook search *
* - encode codebook address *
* - VQ of pitch and codebook gains *
* - find synthesis speech *
* - update states of weighting filter *
*------------------------------------------------------------------------*/
A = A_t; /* pointer to interpolated LPC parameters */
Aq = Aq_t; /* pointer to interpolated quantized LPC parameters */
evenSubfr = 0; move16 ();
subfrNr = -1; move16 ();
for (i_subfr = 0; i_subfr < L_FRAME; i_subfr += L_SUBFR)
{
subfrNr = add(subfrNr, 1);
evenSubfr = sub(1, evenSubfr);
/* Save states for the MR475 mode */
test(); test();
if ((evenSubfr != 0) && (sub(*usedMode, MR475) == 0))
{
Copy(st->mem_syn, mem_syn_save, M);
Copy(st->mem_w0, mem_w0_save, M);
Copy(st->mem_err, mem_err_save, M);
sharp_save = st->sharp;
}
/*-----------------------------------------------------------------*
* - Preprocessing of subframe *
*-----------------------------------------------------------------*/
test();
if (sub(*usedMode, MR475) != 0)
{
subframePreProc(*usedMode, gamma1, gamma1_12k2,
gamma2, A, Aq, &st->speech[i_subfr],
st->mem_err, st->mem_w0, st->zero,
st->ai_zero, &st->exc[i_subfr],
st->h1, xn, res, st->error);
}
else
{ /* MR475 */
subframePreProc(*usedMode, gamma1, gamma1_12k2,
gamma2, A, Aq, &st->speech[i_subfr],
st->mem_err, mem_w0_save, st->zero,
st->ai_zero, &st->exc[i_subfr],
st->h1, xn, res, st->error);
/* save impulse response (modified in cbsearch) */
test ();
if (evenSubfr != 0)
{
Copy (st->h1, h1_sf0, L_SUBFR);
}
}
/* copy the LP residual (res2 is modified in the CL LTP search) */
Copy (res, res2, L_SUBFR);
fwc (); /* function worst case */
/*-----------------------------------------------------------------*
* - Closed-loop LTP search *
*-----------------------------------------------------------------*/
cl_ltp(st->clLtpSt, st->tonStabSt, *usedMode, i_subfr, T_op, st->h1,
&st->exc[i_subfr], res2, xn, lsp_flag, xn2, y1,
&T0, &T0_frac, &gain_pit, gCoeff, &ana,
&gp_limit);
/* update LTP lag history */
move16 (); test(); test ();
if ((subfrNr == 0) && (st->ol_gain_flg[0] > 0))
{
st->old_lags[1] = T0; move16 ();
}
move16 (); test(); test ();
if ((sub(subfrNr, 3) == 0) && (st->ol_gain_flg[1] > 0))
{
st->old_lags[0] = T0; move16 ();
}
fwc (); /* function worst case */
/*-----------------------------------------------------------------*
* - Inovative codebook search (find index and gain) *
*-----------------------------------------------------------------*/
cbsearch(xn2, st->h1, T0, st->sharp, gain_pit, res2,
code, y2, &ana, *usedMode, subfrNr);
fwc (); /* function worst case */
/*------------------------------------------------------*
* - Quantization of gains. *
*------------------------------------------------------*/
gainQuant(st->gainQuantSt, *usedMode, res, &st->exc[i_subfr], code,
xn, xn2, y1, y2, gCoeff, evenSubfr, gp_limit,
&gain_pit_sf0, &gain_code_sf0,
&gain_pit, &gain_code, &ana);
fwc (); /* function worst case */
/* update gain history */
update_gp_clipping(st->tonStabSt, gain_pit);
test();
if (sub(*usedMode, MR475) != 0)
{
/* Subframe Post Porcessing */
subframePostProc(st->speech, *usedMode, i_subfr, gain_pit,
gain_code, Aq, synth, xn, code, y1, y2, st->mem_syn,
st->mem_err, st->mem_w0, st->exc, &st->sharp);
}
else
{
test();
if (evenSubfr != 0)
{
i_subfr_sf0 = i_subfr; move16 ();
Copy(xn, xn_sf0, L_SUBFR);
Copy(y2, y2_sf0, L_SUBFR);
Copy(code, code_sf0, L_SUBFR);
T0_sf0 = T0; move16 ();
T0_frac_sf0 = T0_frac; move16 ();
/* Subframe Post Porcessing */
subframePostProc(st->speech, *usedMode, i_subfr, gain_pit,
gain_code, Aq, synth, xn, code, y1, y2,
mem_syn_save, st->mem_err, mem_w0_save,
st->exc, &st->sharp);
st->sharp = sharp_save; move16();
}
else
{
/* update both subframes for the MR475 */
/* Restore states for the MR475 mode */
Copy(mem_err_save, st->mem_err, M);
/* re-build excitation for sf 0 */
Pred_lt_3or6(&st->exc[i_subfr_sf0], T0_sf0, T0_frac_sf0,
L_SUBFR, 1);
Convolve(&st->exc[i_subfr_sf0], h1_sf0, y1, L_SUBFR);
Aq -= MP1;
subframePostProc(st->speech, *usedMode, i_subfr_sf0,
gain_pit_sf0, gain_code_sf0, Aq,
synth, xn_sf0, code_sf0, y1, y2_sf0,
st->mem_syn, st->mem_err, st->mem_w0, st->exc,
&sharp_save); /* overwrites sharp_save */
Aq += MP1;
/* re-run pre-processing to get xn right (needed by postproc) */
/* (this also reconstructs the unsharpened h1 for sf 1) */
subframePreProc(*usedMode, gamma1, gamma1_12k2,
gamma2, A, Aq, &st->speech[i_subfr],
st->mem_err, st->mem_w0, st->zero,
st->ai_zero, &st->exc[i_subfr],
st->h1, xn, res, st->error);
/* re-build excitation sf 1 (changed if lag < L_SUBFR) */
Pred_lt_3or6(&st->exc[i_subfr], T0, T0_frac, L_SUBFR, 1);
Convolve(&st->exc[i_subfr], st->h1, y1, L_SUBFR);
subframePostProc(st->speech, *usedMode, i_subfr, gain_pit,
gain_code, Aq, synth, xn, code, y1, y2,
st->mem_syn, st->mem_err, st->mem_w0,
st->exc, &st->sharp);
}
}
fwc (); /* function worst case */
A += MP1; /* interpolated LPC parameters for next subframe */
Aq += MP1;
}
Copy(&st->old_exc[L_FRAME], &st->old_exc[0], PIT_MAX + L_INTERPOL);
the_end:
/*--------------------------------------------------*
* Update signal for next frame. *
*--------------------------------------------------*/
Copy(&st->old_wsp[L_FRAME], &st->old_wsp[0], PIT_MAX);
Copy(&st->old_speech[L_FRAME], &st->old_speech[0], L_TOTAL - L_FRAME);
fwc (); /* function worst case */
return 0;
}
bool PixPair::Load(
const char *apath,
const char *bpath,
int order,
int bDoG,
int r1,
int r2,
FILE* flog )
{
printf( "\n---- Image loading ----\n" );
clock_t t0 = StartTiming();
/* ----------------------------- */
/* Load and sanity check rasters */
/* ----------------------------- */
uint8 *aras, *bras;
uint32 wa, ha, wb, hb;
int ok = false;
aras = Raster8FromAny( apath, wa, ha, flog );
bras = Raster8FromAny( bpath, wb, hb, flog );
if( !aras || !bras ) {
fprintf( flog,
"PixPair: Picture load failure.\n" );
goto exit;
}
if( wa != wb || ha != hb ) {
fprintf( flog,
"PixPair: Nonmatching picture dimensions.\n" );
goto exit;
}
ok = true;
wf = wa;
hf = ha;
ws = wa;
hs = ha;
scl = 1;
/* -------------- */
/* Resin removal? */
/* -------------- */
//if( dbgCor ) {
// ZeroResin( "bloba.tif", aras );
// ZeroResin( "blobb.tif", bras );
//}
//else {
// ZeroResin( NULL, aras );
// ZeroResin( NULL, bras );
//}
//StopTiming( flog, "Resin removal", t0 );
/* ------- */
/* Flatten */
/* ------- */
LegPolyFlatten( _avf, aras, wf, hf, order );
RasterFree( aras );
LegPolyFlatten( _bvf, bras, wf, hf, order );
RasterFree( bras );
avs_vfy = avs_aln = avf_vfy = avf_aln = &_avf;
bvs_vfy = bvs_aln = bvf_vfy = bvf_aln = &_bvf;
/* ------------- */
/* Apply filters */
/* ------------- */
//{
// vector<CD> kfft;
// double K[] = {
// 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
// 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
// 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
// 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
// 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
// 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
// 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
// 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
// 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
// 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
// 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1};
// Convolve( _avf, _avf, wf, hf, K, 11, 11, true, true, kfft );
// Normalize( _avf );
// Convolve( _bvf, _bvf, wf, hf, K, 11, 11, true, true, kfft );
// Normalize( _bvf );
//}
#if 0
{
vector<CD> kfft;
double K[] = {
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1};
Convolve( _avfflt, _avf, wf, hf, K, 11, 11, true, true, kfft );
Normalize( _avfflt );
Convolve( _bvfflt, _bvf, wf, hf, K, 11, 11, true, true, kfft );
Normalize( _bvfflt );
avs_aln = avf_aln = &_avfflt;
bvs_aln = bvf_aln = &_bvfflt;
bDoG = true;
}
#endif
if( bDoG ) {
vector<double> DoG;
vector<CD> kfft;
int dim = MakeDoGKernel( DoG, r1, r2, flog );
Convolve( _avfflt, _avf, wf, hf,
&DoG[0], dim, dim, true, true, kfft );
Normalize( _avfflt );
Convolve( _bvfflt, _bvf, wf, hf,
&DoG[0], dim, dim, true, true, kfft );
Normalize( _bvfflt );
avs_aln = avf_aln = &_avfflt;
bvs_aln = bvf_aln = &_bvfflt;
}
/* --------------------- */
/* Downsample all images */
/* --------------------- */
if( ws > 2048 || hs >= 2048 ) {
do {
ws /= 2;
hs /= 2;
scl *= 2;
} while( ws > 2048 || hs > 2048 );
fprintf( flog, "PixPair: Scaling by %d\n", scl );
if( ws * scl != wf || hs * scl != hf ) {
fprintf( flog,
"PixPair: Dimensions not multiple of scale!\n" );
goto exit;
}
Downsample( _avs, _avf );
Downsample( _bvs, _bvf );
avs_vfy = avs_aln = &_avs;
bvs_vfy = bvs_aln = &_bvs;
if( bDoG ) {
if( _avfflt.size() ) {
Downsample( _avsflt, _avfflt );
avs_aln = &_avsflt;
}
if( _bvfflt.size() ) {
Downsample( _bvsflt, _bvfflt );
bvs_aln = &_bvsflt;
}
}
}
else
fprintf( flog, "PixPair: Using image scale=1.\n" );
/* ------------------------------ */
/* Write DoG images for debugging */
/* ------------------------------ */
#if 0
if( bDoG ) {
VectorDblToTif8( "DoGa.tif", avs_aln, ws, hs );
VectorDblToTif8( "DoGb.tif", bvs_aln, ws, hs );
}
#endif
/* -------- */
/* Clean up */
/* -------- */
exit:
if( aras )
RasterFree( aras );
if( bras )
RasterFree( bras );
StopTiming( flog, "Image conditioning", t0 );
return ok;
}
Feature
HOGFeatureExtractor::operator()(const CByteImage& img_) const
{
/******** BEGIN TODO ********/
// Compute the Histogram of Oriented Gradients feature
// Steps are:
// 1) Compute gradients in x and y directions. We provide the
// derivative kernel proposed in the paper in _kernelDx and
// _kernelDy.
// 2) Compute gradient magnitude and orientation
// 3) Add contribution each pixel to HOG cells whose
// support overlaps with pixel. Each cell has a support of size
// _cellSize and each histogram has _nAngularBins.
// 4) Normalize HOG for each cell. One simple strategy that is
// is also used in the SIFT descriptor is to first threshold
// the bin values so that no bin value is larger than some
// threshold (we leave it up to you do find this value) and
// then re-normalize the histogram so that it has norm 1. A more
// elaborate normalization scheme is proposed in Dalal & Triggs
// paper but we leave that as extra credit.
//
// Useful functions:
// convertRGB2GrayImage, TypeConvert, WarpGlobal, Convolve
int xCells = ceil(1.*img_.Shape().width / _cellSize);
int yCells = ceil(1.*img_.Shape().height / _cellSize);
CFloatImage HOGHist(xCells, yCells, _nAngularBins);
HOGHist.ClearPixels();
CByteImage gray(img_.Shape());
CFloatImage grayF(img_.Shape().width, img_.Shape().height, 1);
convertRGB2GrayImage(img_, gray);
TypeConvert(gray, grayF);
CFloatImage diffX( img_.Shape()), diffY( img_.Shape());
Convolve(grayF, diffX, _kernelDx);
Convolve(grayF, diffY, _kernelDy);
CFloatImage grad(grayF.Shape()), grad2(grayF.Shape());
CFloatImage angl(grayF.Shape()), angl2(grayF.Shape());
for (int y = 0; y <grayF.Shape().height; y++){
for (int x = 0; x<grayF.Shape().width; x++) {
grad2.Pixel(x,y,0) = (diffX.Pixel(x,y,0) * diffX.Pixel(x,y,0) +
diffY.Pixel(x,y,0) * diffY.Pixel(x,y,0));
angl2.Pixel(x,y,0) = atan(diffY.Pixel(x,y,0) / abs(diffY.Pixel(x,y,0)));
}
}
// Bilinear Filter
ConvolveSeparable(grad2, grad, ConvolveKernel_121,ConvolveKernel_121,1);
ConvolveSeparable(angl2, angl, ConvolveKernel_121,ConvolveKernel_121,1);
//WriteFile(diffX, "angle.tga");
//WriteFile(diffY, "angleG.tga");
for (int y = 0; y <grayF.Shape().height; y++){
for (int x = 0; x<grayF.Shape().width; x++) {
// Fit in the bins
int a = angl.Pixel(x,y,0) / 3.14 * (_nAngularBins) + _nAngularBins/2;
// Histogram
HOGHist.Pixel(floor(1.*x / _cellSize),
floor(1.*y / _cellSize),
a) += grad.Pixel(x,y,0);
}
}
// Normalization
float threshold = 0.7;
for (int y = 0; y < yCells; y++){
for (int x = 0; x < xCells; x++){
float total = 0;
for (int a = 0; a < _nAngularBins; a++) {
if (HOGHist.Pixel(x,y,a) > threshold)
HOGHist.Pixel(x,y,a) = threshold;
// Sum for normalization
total += HOGHist.Pixel(x,y,a);
}
for (int a = 0;a< _nAngularBins; a++) {
HOGHist.Pixel(x,y,a) /= total;
}
}
}
return HOGHist;
/******** END TODO ********/
}
CFloatImage
SupportVectorMachine::predictSlidingWindow(const Feature& feat) const
{
CFloatImage score(CShape(feat.Shape().width,feat.Shape().height,1));
score.ClearPixels();
/******** BEGIN TODO ********/
// Sliding window prediction.
//
// In this project we are using a linear SVM. This means that
// it's classification function is very simple, consisting of a
// dot product of the feature vector with a set of weights learned
// during training, followed by a subtraction of a bias term
//
// pred <- dot(feat, weights) - bias term
//
// Now this is very simple to compute when we are dealing with
// cropped images, our computed features have the same dimensions
// as the SVM weights. Things get a little more tricky when you
// want to evaluate this function over all possible subwindows of
// a larger feature, one that we would get by running our feature
// extraction on an entire image.
//
// Here you will evaluate the above expression by breaking
// the dot product into a series of convolutions (remember that
// a convolution can be though of as a point wise dot product with
// the convolution kernel), each one with a different band.
//
// Convolve each band of the SVM weights with the corresponding
// band in feat, and add the resulting score image. The final
// step is to subtract the SVM bias term given by this->getBiasTerm().
//
// Hint: you might need to set the origin for the convolution kernel
// in order to get the result from convoltion to be correctly centered
//
// Useful functions:
// Convolve, BandSelect, this->getWeights(), this->getBiasTerm()
//printf("TODO: SupportVectorMachine.cpp:273\n"); exit(EXIT_FAILURE);
Feature weights = getWeights();
for (int b=0; b<feat.Shape().nBands; b++){
CFloatImage currentBandWeights = CFloatImage(weights.Shape().width, weights.Shape().height, 1);
CFloatImage currentBandFeatures = CFloatImage(feat.Shape().width, feat.Shape().height, 1);
CFloatImage convolved = CFloatImage(CShape(feat.Shape().width, feat.Shape().height, 1));
CFloatImage final(CShape(feat.Shape().width, feat.Shape().height, 1));
BandSelect(weights, currentBandWeights, b, 0);
BandSelect(feat, currentBandFeatures, b, 0);
currentBandWeights.origin[0] = weights.origin[0];
currentBandWeights.origin[1] = weights.origin[1];
Convolve(feat, convolved, currentBandWeights);
BandSelect(convolved, final, b, 0);
try{
score += final;
} catch (CError err) {
printf("OH NOES: the final chapter!");
}
}
score-=getBiasTerm();
/******** END TODO ********/
return score;
}
void ConvolveSeparable(CImageOf<T> src, CImageOf<T>& dst,
CFloatImage x_kernel, CFloatImage y_kernel,
float scale, float offset,
int decimate, int interpolate)
{
// Allocate the result, if necessary
CShape dShape = src.Shape();
if (decimate > 1)
{
dShape.width = (dShape.width + decimate-1) / decimate;
dShape.height = (dShape.height + decimate-1) / decimate;
}
dst.ReAllocate(dShape, false);
// Allocate the intermediate images
CImageOf<T> tmpImg1(src.Shape());
//CImageOf<T> tmpImgCUDA(src.Shape());
CImageOf<T> tmpImg2(src.Shape());
// Create a proper vertical convolution kernel
CFloatImage v_kernel(1, y_kernel.Shape().width, 1);
for (int k = 0; k < y_kernel.Shape().width; k++)
v_kernel.Pixel(0, k, 0) = y_kernel.Pixel(k, 0, 0);
v_kernel.origin[1] = y_kernel.origin[0];
#ifdef RUN_ON_GPU
// Modifications for integrating CUDA kernels
BinomialFilterType type;
profilingTimer->startTimer();
// CUDA Convolve
switch (x_kernel.Shape().width)
{
case 3:
type = BINOMIAL6126;
break;
case 5:
type = BINOMIAL14641;
break;
default:
// Unsupported kernel case
throw CError("Convolution kernel Unknown");
assert(false);
}
// Skip copy if decimation is not required
if (decimate != 1) CudaConvolveXY(src, tmpImg2, type);
else CudaConvolveXY(src, dst, type);
printf("\nGPU convolution time = %f ms\n", profilingTimer->stopAndGetTimerValue());
#else
profilingTimer->startTimer();
//VerifyComputedData(&tmpImg2.Pixel(0, 0, 0), &tmpImgCUDA.Pixel(0, 0, 0), 7003904);
// Perform the two convolutions
Convolve(src, tmpImg1, x_kernel, 1.0f, 0.0f);
Convolve(tmpImg1, tmpImg2, v_kernel, scale, offset);
printf("\nCPU Convolution time = %f ms\n", profilingTimer->stopAndGetTimerValue());
#endif
profilingTimer->startTimer();
// Downsample or copy
// Skip decimate and recopy if not required
#ifdef RUN_ON_GPU
if (decimate != 1)
{
#endif
for (int y = 0; y < dShape.height; y++)
{
T* sPtr = &tmpImg2.Pixel(0, y * decimate, 0);
T* dPtr = &dst.Pixel(0, y, 0);
int nB = dShape.nBands;
for (int x = 0; x < dShape.width; x++)
{
for (int b = 0; b < nB; b++)
dPtr[b] = sPtr[b];
sPtr += decimate * nB;
dPtr += nB;
}
}
#ifdef RUN_ON_GPU
}
#endif
printf("\nDecimate/Recopy took = %f ms\n", profilingTimer->stopAndGetTimerValue());
}
void Coder_ld8h(
Word16 ana[], /* (o) : analysis parameters */
Word16 rate /* input : rate selector/frame =0 6.4kbps , =1 8kbps,= 2 11.8 kbps*/
)
{
/* LPC analysis */
Word16 r_l_fwd[MP1], r_h_fwd[MP1]; /* Autocorrelations low and hi (forward) */
Word32 r_bwd[M_BWDP1]; /* Autocorrelations (backward) */
Word16 r_l_bwd[M_BWDP1]; /* Autocorrelations low (backward) */
Word16 r_h_bwd[M_BWDP1]; /* Autocorrelations high (backward) */
Word16 rc_fwd[M]; /* Reflection coefficients : forward analysis */
Word16 rc_bwd[M_BWD]; /* Reflection coefficients : backward analysis */
Word16 A_t_fwd[MP1*2]; /* A(z) forward unquantized for the 2 subframes */
Word16 A_t_fwd_q[MP1*2]; /* A(z) forward quantized for the 2 subframes */
Word16 A_t_bwd[2*M_BWDP1]; /* A(z) backward for the 2 subframes */
Word16 *Aq; /* A(z) "quantized" for the 2 subframes */
Word16 *Ap; /* A(z) "unquantized" for the 2 subframes */
Word16 *pAp, *pAq;
Word16 Ap1[M_BWDP1]; /* A(z) with spectral expansion */
Word16 Ap2[M_BWDP1]; /* A(z) with spectral expansion */
Word16 lsp_new[M], lsp_new_q[M]; /* LSPs at 2th subframe */
Word16 lsf_int[M]; /* Interpolated LSF 1st subframe. */
Word16 lsf_new[M];
Word16 lp_mode; /* Backward / Forward Indication mode */
Word16 m_ap, m_aq, i_gamma;
Word16 code_lsp[2];
/* Other vectors */
Word16 h1[L_SUBFR]; /* Impulse response h1[] */
Word16 xn[L_SUBFR]; /* Target vector for pitch search */
Word16 xn2[L_SUBFR]; /* Target vector for codebook search */
Word16 code[L_SUBFR]; /* Fixed codebook excitation */
Word16 y1[L_SUBFR]; /* Filtered adaptive excitation */
Word16 y2[L_SUBFR]; /* Filtered fixed codebook excitation */
Word16 g_coeff[4]; /* Correlations between xn & y1 */
Word16 res2[L_SUBFR]; /* residual after long term prediction*/
Word16 g_coeff_cs[5];
Word16 exp_g_coeff_cs[5]; /* Correlations between xn, y1, & y2
<y1,y1>, -2<xn,y1>,
<y2,y2>, -2<xn,y2>, 2<y1,y2> */
/* Scalars */
Word16 i, j, k, i_subfr;
Word16 T_op, T0, T0_min, T0_max, T0_frac;
Word16 gain_pit, gain_code, index;
Word16 taming, pit_sharp;
Word16 sat_filter;
Word32 L_temp;
Word16 freq_cur[M];
Word16 temp;
/*------------------------------------------------------------------------*
* - Perform LPC analysis: *
* * autocorrelation + lag windowing *
* * Levinson-durbin algorithm to find a[] *
* * convert a[] to lsp[] *
* * quantize and code the LSPs *
* * find the interpolated LSPs and convert to a[] for the 2 *
* subframes (both quantized and unquantized) *
*------------------------------------------------------------------------*/
/* ------------------- */
/* LP Forward analysis */
/* ------------------- */
Autocorr(p_window, M, r_h_fwd, r_l_fwd); /* Autocorrelations */
Lag_window(M, r_h_fwd, r_l_fwd); /* Lag windowing */
Levinsone(M, r_h_fwd, r_l_fwd, &A_t_fwd[MP1], rc_fwd, old_A_fwd, old_rc_fwd); /* Levinson Durbin */
Az_lsp(&A_t_fwd[MP1], lsp_new, lsp_old); /* From A(z) to lsp */
/* -------------------- */
/* LP Backward analysis */
/* -------------------- */
/* -------------------- */
/* LP Backward analysis */
/* -------------------- */
if ( rate== G729E) {
/* LPC recursive Window as in G728 */
autocorr_hyb_window(synth, r_bwd, rexp); /* Autocorrelations */
Lag_window_bwd(r_bwd, r_h_bwd, r_l_bwd); /* Lag windowing */
/* Fixed Point Levinson (as in G729) */
Levinsone(M_BWD, r_h_bwd, r_l_bwd, &A_t_bwd[M_BWDP1], rc_bwd,
old_A_bwd, old_rc_bwd);
/* Tests saturation of A_t_bwd */
sat_filter = 0;
for (i=M_BWDP1; i<2*M_BWDP1; i++) if (A_t_bwd[i] >= 32767) sat_filter = 1;
if (sat_filter == 1) Copy(A_t_bwd_mem, &A_t_bwd[M_BWDP1], M_BWDP1);
else Copy(&A_t_bwd[M_BWDP1], A_t_bwd_mem, M_BWDP1);
/* Additional bandwidth expansion on backward filter */
Weight_Az(&A_t_bwd[M_BWDP1], GAMMA_BWD, M_BWD, &A_t_bwd[M_BWDP1]);
}
/*--------------------------------------------------*
* Update synthesis signal for next frame. *
*--------------------------------------------------*/
Copy(&synth[L_FRAME], &synth[0], MEM_SYN_BWD);
/*--------------------------------------------------------------------*
* Find interpolated LPC parameters in all subframes (unquantized). *
* The interpolated parameters are in array A_t[] of size (M+1)*4 *
*--------------------------------------------------------------------*/
if( prev_lp_mode == 0) {
Int_lpc(lsp_old, lsp_new, lsf_int, lsf_new, A_t_fwd);
}
else {
/* no interpolation */
/* unquantized */
Lsp_Az(lsp_new, A_t_fwd); /* Subframe 1 */
Lsp_lsf(lsp_new, lsf_new, M); /* transformation from LSP to LSF (freq.domain) */
Copy(lsf_new, lsf_int, M); /* Subframe 1 */
}
/* ---------------- */
/* LSP quantization */
/* ---------------- */
Qua_lspe(lsp_new, lsp_new_q, code_lsp, freq_prev, freq_cur);
/*--------------------------------------------------------------------*
* Find interpolated LPC parameters in all subframes (quantized) *
* the quantized interpolated parameters are in array Aq_t[] *
*--------------------------------------------------------------------*/
if( prev_lp_mode == 0) {
Int_qlpc(lsp_old_q, lsp_new_q, A_t_fwd_q);
}
else {
/* no interpolation */
Lsp_Az(lsp_new_q, &A_t_fwd_q[MP1]); /* Subframe 2 */
Copy(&A_t_fwd_q[MP1], A_t_fwd_q, MP1); /* Subframe 1 */
}
/*---------------------------------------------------------------------*
* - Decision for the switch Forward / Backward *
*---------------------------------------------------------------------*/
if(rate == G729E) {
set_lpc_modeg(speech, A_t_fwd_q, A_t_bwd, &lp_mode,
lsp_new, lsp_old, &bwd_dominant, prev_lp_mode, prev_filter,
&C_int, &glob_stat, &stat_bwd, &val_stat_bwd);
}
else {
update_bwd( &lp_mode, &bwd_dominant, &C_int, &glob_stat);
}
/* ---------------------------------- */
/* update the LSPs for the next frame */
/* ---------------------------------- */
Copy(lsp_new, lsp_old, M);
/*----------------------------------------------------------------------*
* - Find the weighted input speech w_sp[] for the whole speech frame *
*----------------------------------------------------------------------*/
if(lp_mode == 0) {
m_ap = M;
if (bwd_dominant == 0) Ap = A_t_fwd;
else Ap = A_t_fwd_q;
perc_var(gamma1, gamma2, lsf_int, lsf_new, rc_fwd);
}
else {
if (bwd_dominant == 0) {
m_ap = M;
Ap = A_t_fwd;
}
else {
m_ap = M_BWD;
Ap = A_t_bwd;
}
perc_vare(gamma1, gamma2, bwd_dominant);
}
pAp = Ap;
for (i=0; i<2; i++) {
Weight_Az(pAp, gamma1[i], m_ap, Ap1);
Weight_Az(pAp, gamma2[i], m_ap, Ap2);
Residue(m_ap, Ap1, &speech[i*L_SUBFR], &wsp[i*L_SUBFR], L_SUBFR);
Syn_filte(m_ap, Ap2, &wsp[i*L_SUBFR], &wsp[i*L_SUBFR], L_SUBFR,
&mem_w[M_BWD-m_ap], 0);
for(j=0; j<M_BWD; j++) mem_w[j] = wsp[i*L_SUBFR+L_SUBFR-M_BWD+j];
pAp += m_ap+1;
}
*ana++ = rate+ (Word16)2; /* bit rate mode */
if(lp_mode == 0) {
m_aq = M;
Aq = A_t_fwd_q;
/* update previous filter for next frame */
Copy(&Aq[MP1], prev_filter, MP1);
for(i=MP1; i <M_BWDP1; i++) prev_filter[i] = 0;
for(j=MP1; j<M_BWDP1; j++) ai_zero[j] = 0;
}
else {
m_aq = M_BWD;
Aq = A_t_bwd;
if (bwd_dominant == 0) {
for(j=MP1; j<M_BWDP1; j++) ai_zero[j] = 0;
}
/* update previous filter for next frame */
Copy(&Aq[M_BWDP1], prev_filter, M_BWDP1);
}
if (rate == G729E) *ana++ = lp_mode;
/*----------------------------------------------------------------------*
* - Find the weighted input speech w_sp[] for the whole speech frame *
* - Find the open-loop pitch delay *
*----------------------------------------------------------------------*/
if( lp_mode == 0) {
Copy(lsp_new_q, lsp_old_q, M);
Lsp_prev_update(freq_cur, freq_prev);
*ana++ = code_lsp[0];
*ana++ = code_lsp[1];
}
/* Find open loop pitch lag */
T_op = Pitch_ol(wsp, PIT_MIN, PIT_MAX, L_FRAME);
/* Range for closed loop pitch search in 1st subframe */
T0_min = sub(T_op, 3);
if (sub(T0_min,PIT_MIN)<0) {
T0_min = PIT_MIN;
}
T0_max = add(T0_min, 6);
if (sub(T0_max ,PIT_MAX)>0)
{
T0_max = PIT_MAX;
T0_min = sub(T0_max, 6);
}
/*------------------------------------------------------------------------*
* Loop for every subframe in the analysis frame *
*------------------------------------------------------------------------*
* To find the pitch and innovation parameters. The subframe size is *
* L_SUBFR and the loop is repeated 2 times. *
* - find the weighted LPC coefficients *
* - find the LPC residual signal res[] *
* - compute the target signal for pitch search *
* - compute impulse response of weighted synthesis filter (h1[]) *
* - find the closed-loop pitch parameters *
* - encode the pitch delay *
* - update the impulse response h1[] by including fixed-gain pitch *
* - find target vector for codebook search *
* - codebook search *
* - encode codebook address *
* - VQ of pitch and codebook gains *
* - find synthesis speech *
* - update states of weighting filter *
*------------------------------------------------------------------------*/
pAp = Ap; /* pointer to interpolated "unquantized"LPC parameters */
pAq = Aq; /* pointer to interpolated "quantized" LPC parameters */
i_gamma = 0;
for (i_subfr = 0; i_subfr < L_FRAME; i_subfr += L_SUBFR) {
/*---------------------------------------------------------------*
* Find the weighted LPC coefficients for the weighting filter. *
*---------------------------------------------------------------*/
Weight_Az(pAp, gamma1[i_gamma], m_ap, Ap1);
Weight_Az(pAp, gamma2[i_gamma], m_ap, Ap2);
/*---------------------------------------------------------------*
* Compute impulse response, h1[], of weighted synthesis filter *
*---------------------------------------------------------------*/
for (i = 0; i <=m_ap; i++) ai_zero[i] = Ap1[i];
Syn_filte(m_aq, pAq, ai_zero, h1, L_SUBFR, zero, 0);
Syn_filte(m_ap, Ap2, h1, h1, L_SUBFR, zero, 0);
/*------------------------------------------------------------------------*
* *
* Find the target vector for pitch search: *
* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ *
* *
* |------| res[n] *
* speech[n]---| A(z) |-------- *
* |------| | |--------| error[n] |------| *
* zero -- (-)--| 1/A(z) |-----------| W(z) |-- target *
* exc |--------| |------| *
* *
* Instead of subtracting the zero-input response of filters from *
* the weighted input speech, the above configuration is used to *
* compute the target vector. This configuration gives better performance *
* with fixed-point implementation. The memory of 1/A(z) is updated by *
* filtering (res[n]-exc[n]) through 1/A(z), or simply by subtracting *
* the synthesis speech from the input speech: *
* error[n] = speech[n] - syn[n]. *
* The memory of W(z) is updated by filtering error[n] through W(z), *
* or more simply by subtracting the filtered adaptive and fixed *
* codebook excitations from the target: *
* target[n] - gain_pit*y1[n] - gain_code*y2[n] *
* as these signals are already available. *
* *
*------------------------------------------------------------------------*/
Residue(m_aq, pAq, &speech[i_subfr], &exc[i_subfr], L_SUBFR); /* LPC residual */
for (i=0; i<L_SUBFR; i++) res2[i] = exc[i_subfr+i];
Syn_filte(m_aq, pAq, &exc[i_subfr], error, L_SUBFR,
&mem_err[M_BWD-m_aq], 0);
Residue(m_ap, Ap1, error, xn, L_SUBFR);
Syn_filte(m_ap, Ap2, xn, xn, L_SUBFR, &mem_w0[M_BWD-m_ap], 0); /* target signal xn[]*/
/*----------------------------------------------------------------------*
* Closed-loop fractional pitch search *
*----------------------------------------------------------------------*/
T0 = Pitch_fr3cp(&exc[i_subfr], xn, h1, L_SUBFR, T0_min, T0_max,
i_subfr, &T0_frac, rate);
index = Enc_lag3cp(T0, T0_frac, &T0_min, &T0_max,PIT_MIN,PIT_MAX,
i_subfr, rate);
*ana++ = index;
if ( (i_subfr == 0) && (rate != G729D) ) {
*ana = Parity_Pitch(index);
if( rate == G729E) {
*ana ^= (shr(index, 1) & 0x0001);
}
ana++;
}
/*-----------------------------------------------------------------*
* - find unity gain pitch excitation (adaptive codebook entry) *
* with fractional interpolation. *
* - find filtered pitch exc. y1[]=exc[] convolve with h1[]) *
* - compute pitch gain and limit between 0 and 1.2 *
* - update target vector for codebook search *
* - find LTP residual. *
*-----------------------------------------------------------------*/
Pred_lt_3(&exc[i_subfr], T0, T0_frac, L_SUBFR);
Convolve(&exc[i_subfr], h1, y1, L_SUBFR);
gain_pit = G_pitch(xn, y1, g_coeff, L_SUBFR);
/* clip pitch gain if taming is necessary */
taming = test_err(T0, T0_frac);
if( taming == 1){
if (sub(gain_pit, GPCLIP) > 0) {
gain_pit = GPCLIP;
}
}
/* xn2[i] = xn[i] - y1[i] * gain_pit */
for (i = 0; i < L_SUBFR; i++) {
L_temp = L_mult(y1[i], gain_pit);
L_temp = L_shl(L_temp, 1); /* gain_pit in Q14 */
xn2[i] = sub(xn[i], extract_h(L_temp));
}
/*-----------------------------------------------------*
* - Innovative codebook search. *
*-----------------------------------------------------*/
switch (rate) {
case G729: /* 8 kbit/s */
{
/* case 8 kbit/s */
index = ACELP_Codebook(xn2, h1, T0, sharp, i_subfr, code, y2, &i);
*ana++ = index; /* Positions index */
*ana++ = i; /* Signs index */
break;
}
case G729D: /* 6.4 kbit/s */
{
index = ACELP_CodebookD(xn2, h1, T0, sharp, code, y2, &i);
*ana++ = index; /* Positions index */
*ana++ = i; /* Signs index */
break;
}
case G729E: /* 11.8 kbit/s */
{
/*-----------------------------------------------------------------*
* Include fixed-gain pitch contribution into impulse resp. h[] *
*-----------------------------------------------------------------*/
pit_sharp = shl(sharp, 1); /* From Q14 to Q15 */
if(T0 < L_SUBFR) {
for (i = T0; i < L_SUBFR; i++){ /* h[i] += pitch_sharp*h[i-T0] */
h1[i] = add(h1[i], mult(h1[i-T0], pit_sharp));
}
}
/* calculate residual after long term prediction */
/* res2[i] -= exc[i+i_subfr] * gain_pit */
for (i = 0; i < L_SUBFR; i++) {
L_temp = L_mult(exc[i+i_subfr], gain_pit);
L_temp = L_shl(L_temp, 1); /* gain_pit in Q14 */
res2[i] = sub(res2[i], extract_h(L_temp));
}
if (lp_mode == 0) ACELP_10i40_35bits(xn2, res2, h1, code, y2, ana); /* Forward */
else ACELP_12i40_44bits(xn2, res2, h1, code, y2, ana); /* Backward */
ana += 5;
/*-----------------------------------------------------------------*
* Include fixed-gain pitch contribution into code[]. *
*-----------------------------------------------------------------*/
if(T0 < L_SUBFR) {
for (i = T0; i < L_SUBFR; i++) { /* code[i] += pitch_sharp*code[i-T0] */
code[i] = add(code[i], mult(code[i-T0], pit_sharp));
}
}
break;
}
default : {
printf("Unrecognized bit rate\n");
exit(-1);
}
} /* end of switch */
/*-----------------------------------------------------*
* - Quantization of gains. *
*-----------------------------------------------------*/
g_coeff_cs[0] = g_coeff[0]; /* <y1,y1> */
exp_g_coeff_cs[0] = negate(g_coeff[1]); /* Q-Format:XXX -> JPN */
g_coeff_cs[1] = negate(g_coeff[2]); /* (xn,y1) -> -2<xn,y1> */
exp_g_coeff_cs[1] = negate(add(g_coeff[3], 1)); /* Q-Format:XXX -> JPN */
Corr_xy2( xn, y1, y2, g_coeff_cs, exp_g_coeff_cs ); /* Q0 Q0 Q12 ^Qx ^Q0 */
/* g_coeff_cs[3]:exp_g_coeff_cs[3] = <y2,y2> */
/* g_coeff_cs[4]:exp_g_coeff_cs[4] = -2<xn,y2> */
/* g_coeff_cs[5]:exp_g_coeff_cs[5] = 2<y1,y2> */
if (rate == G729D)
index = Qua_gain_6k(code, g_coeff_cs, exp_g_coeff_cs, L_SUBFR,
&gain_pit, &gain_code, taming, past_qua_en);
else
index = Qua_gain_8k(code, g_coeff_cs, exp_g_coeff_cs, L_SUBFR,
&gain_pit, &gain_code, taming, past_qua_en);
*ana++ = index;
/*------------------------------------------------------------*
* - Update pitch sharpening "sharp" with quantized gain_pit *
*------------------------------------------------------------*/
sharp = gain_pit;
if (sub(sharp, SHARPMAX) > 0) sharp = SHARPMAX;
else {
if (sub(sharp, SHARPMIN) < 0) sharp = SHARPMIN;
}
/*------------------------------------------------------*
* - Find the total excitation *
* - find synthesis speech corresponding to exc[] *
* - update filters memories for finding the target *
* vector in the next subframe *
* (update error[-m..-1] and mem_w_err[]) *
* update error function for taming process *
*------------------------------------------------------*/
for (i = 0; i < L_SUBFR; i++) {
/* exc[i] = gain_pit*exc[i] + gain_code*code[i]; */
/* exc[i] in Q0 gain_pit in Q14 */
/* code[i] in Q13 gain_cod in Q1 */
L_temp = L_mult(exc[i+i_subfr], gain_pit);
L_temp = L_mac(L_temp, code[i], gain_code);
L_temp = L_shl(L_temp, 1);
exc[i+i_subfr] = round(L_temp);
}
update_exc_err(gain_pit, T0);
Syn_filte(m_aq, pAq, &exc[i_subfr], &synth_ptr[i_subfr], L_SUBFR,
&mem_syn[M_BWD-m_aq], 0);
for(j=0; j<M_BWD; j++) mem_syn[j] = synth_ptr[i_subfr+L_SUBFR-M_BWD+j];
for (i = L_SUBFR-M_BWD, j = 0; i < L_SUBFR; i++, j++) {
mem_err[j] = sub(speech[i_subfr+i], synth_ptr[i_subfr+i]);
temp = extract_h(L_shl( L_mult(y1[i], gain_pit), 1) );
k = extract_h(L_shl( L_mult(y2[i], gain_code), 2) );
mem_w0[j] = sub(xn[i], add(temp, k));
}
pAp += m_ap+1;
pAq += m_aq+1;
i_gamma = add(i_gamma,1);
}
/*--------------------------------------------------*
* Update signal for next frame. *
* -> shift to the left by L_FRAME: *
* speech[], wsp[] and exc[] *
*--------------------------------------------------*/
Copy(&old_speech[L_FRAME], &old_speech[0], L_TOTAL-L_FRAME);
Copy(&old_wsp[L_FRAME], &old_wsp[0], PIT_MAX);
Copy(&old_exc[L_FRAME], &old_exc[0], PIT_MAX+L_INTERPOL);
prev_lp_mode = lp_mode;
return;
}
// Compute MOPs descriptors.
void ComputeMOPSDescriptors(CFloatImage &image, FeatureSet &features)
{
int w = image.Shape().width; // image width
int h = image.Shape().height; // image height
// Create grayscale image used for Harris detection
CFloatImage grayImage=ConvertToGray(image);
// Apply a 7x7 gaussian blur to the grayscale image
CFloatImage blurImage(w,h,1);
Convolve(grayImage, blurImage, ConvolveKernel_7x7);
// Transform matrices
CTransform3x3 xform;
CTransform3x3 trans1;
CTransform3x3 rotate;
CTransform3x3 scale;
CTransform3x3 trans2;
// Declare additional variables
float pxl; // pixel value
double mean, sq_sum, stdev; // variables for normailizing data set
// This image represents the window around the feature you need to compute to store as the feature descriptor
const int windowSize = 8;
CFloatImage destImage(windowSize, windowSize, 1);
for (vector<Feature>::iterator i = features.begin(); i != features.end(); i++) {
Feature &f = *i;
// Compute the transform from each pixel in the 8x8 image to sample from the appropriate
// pixels in the 40x40 rotated window surrounding the feature
trans1 = CTransform3x3::Translation(f.x, f.y); // translate window to feature point
rotate = CTransform3x3::Rotation(f.angleRadians * 180.0 / PI); // rotate window by angle
scale = CTransform3x3::Scale(5.0); // scale window by 5
trans2 = CTransform3x3::Translation(-windowSize/2, -windowSize/2); // translate window to origin
// transform resulting from combining above transforms
xform = trans1*scale*rotate*trans2;
//Call the Warp Global function to do the mapping
WarpGlobal(blurImage, destImage, xform, eWarpInterpLinear);
// Resize data field for a 8x8 square window
f.data.resize(windowSize * windowSize);
// Find mean of window
mean = 0;
for (int y = 0; y < windowSize; y++) {
for (int x = 0; x < windowSize; x++) {
pxl = destImage.Pixel(x, y, 0);
f.data[y*windowSize + x] = pxl;
mean += pxl/(windowSize*windowSize);
}
}
// Find standard deviation of window
sq_sum = 0;
for (int k = 0; k < windowSize*windowSize; k++) {
sq_sum += (mean - f.data[k]) * (mean - f.data[k]);
}
stdev = sqrt(sq_sum/(windowSize*windowSize));
// Normalize window to have 0 mean and unit variance by subtracting
// by mean and dividing by standard deviation
for (int k = 0; k < windowSize*windowSize; k++) {
f.data[k] = (f.data[k]-mean)/stdev;
}
}
}
static void Norm_Corr(Word16 exc[], Word16 xn[], Word16 h[], Word16 L_subfr,
Word16 t_min, Word16 t_max, Word16 corr_norm[])
{
Word16 i,j,k;
Word16 corr_h, corr_l, norm_h, norm_l;
Word32 s, L_temp;
Word16 excf[L_SUBFR];
Word16 scaling, h_fac, *s_excf, scaled_excf[L_SUBFR];
k = negate(t_min);
/* compute the filtered excitation for the first delay t_min */
Convolve(&exc[k], h, excf, L_subfr);
/* scaled "excf[]" to avoid overflow */
for(j=0; j<L_subfr; j++)
scaled_excf[j] = shr(excf[j], 2);
/* Compute energy of excf[] for danger of overflow */
s = 0;
for (j = 0; j < L_subfr; j++)
s = L_mac(s, excf[j], excf[j]);
L_temp = L_sub(s, 67108864L);
if (L_temp <= 0L) /* if (s <= 2^26) */
{
s_excf = excf;
h_fac = 15-12; /* h in Q12 */
scaling = 0;
}
else {
s_excf = scaled_excf; /* "excf[]" is divide by 2 */
h_fac = 15-12-2; /* h in Q12, divide by 2 */
scaling = 2;
}
/* loop for every possible period */
for (i = t_min; i <= t_max; i++)
{
/* Compute 1/sqrt(energy of excf[]) */
s = 0;
for (j = 0; j < L_subfr; j++)
s = L_mac(s, s_excf[j], s_excf[j]);
s = Inv_sqrt(s); /* Result in Q30 */
L_Extract(s, &norm_h, &norm_l);
/* Compute correlation between xn[] and excf[] */
s = 0;
for (j = 0; j < L_subfr; j++)
s = L_mac(s, xn[j], s_excf[j]);
L_Extract(s, &corr_h, &corr_l);
/* Normalize correlation = correlation * (1/sqrt(energy)) */
s = Mpy_32(corr_h, corr_l, norm_h, norm_l);
corr_norm[i] = extract_h(L_shl(s, 16)); /* Result is on 16 bits */
/* modify the filtered excitation excf[] for the next iteration */
if( sub(i, t_max) != 0)
{
k=sub(k,1);
for (j = L_subfr-(Word16)1; j > 0; j--)
{
s = L_mult(exc[k], h[j]);
s = L_shl(s, h_fac); /* h is in Q(12-scaling) */
s_excf[j] = add(extract_h(s), s_excf[j-1]);
}
s_excf[0] = shr(exc[k], scaling);
}
}
return;
}
void
SupportVectorMachine::predictSlidingWindow(const Feature &feat, CFloatImage &response) const
{
response.ReAllocate(CShape(feat.Shape().width, feat.Shape().height, 1));
response.ClearPixels();
/******** BEGIN TODO ********/
// Sliding window prediction.
//
// In this project we are using a linear SVM. This means that
// it's classification function is very simple, consisting of a
// dot product of the feature vector with a set of weights learned
// during training, followed by a subtraction of a bias term
//
// pred <- dot(feat, weights) - bias term
//
// Now this is very simple to compute when we are dealing with
// cropped images, our computed features have the same dimensions
// as the SVM weights. Things get a little more tricky when you
// want to evaluate this function over all possible subwindows of
// a larger feature, one that we would get by running our feature
// extraction on an entire image.
//
// Here you will evaluate the above expression by breaking
// the dot product into a series of convolutions (remember that
// a convolution can be though of as a point wise dot product with
// the convolution kernel), each one with a different band.
//
// Convolve each band of the SVM weights with the corresponding
// band in feat, and add the resulting score image. The final
// step is to subtract the SVM bias term given by this->getBiasTerm().
//
// Hint: you might need to set the origin for the convolution kernel
// in order to get the result from convoltion to be correctly centered
//
// Useful functions:
// Convolve, BandSelect, this->getWeights(), this->getBiasTerm()
Feature weights = this->getWeights();
int nWtBands = weights.Shape().nBands;
// Set the center of the window as the origin for the conv. kernel
for (int band = 0; band < nWtBands; band++)
{
// Select a band
CFloatImage featBand;
CFloatImage weightBand;
BandSelect(feat, featBand, band, 0);
BandSelect(weights, weightBand, band, 0);
// Set the origin of the kernel
weightBand.origin[0] = weights.Shape().width / 2;
weightBand.origin[1] = weights.Shape().height / 2;
// Compute the dot product
CFloatImage dotproduct;
dotproduct.ClearPixels();
Convolve(featBand, dotproduct, weightBand);
// Add the resulting score image
for (int y = 0; y < feat.Shape().height; y++)
{
for (int x = 0; x < feat.Shape().width; x++)
{
response.Pixel(x, y, 0) += dotproduct.Pixel(x, y, 0);
}
// End of x loop
}
// End of y loop
}
// End of band loop
// Substract the SVM bias term
for (int y = 0; y < feat.Shape().height; y++)
{
for (int x = 0; x < feat.Shape().width; x++)
{
response.Pixel(x, y, 0) -= this->getBiasTerm();
}
// End of x loop
}
// End of y loop
/******** END TODO ********/
}
// Compute MOPs descriptors.
void ComputeMOPSDescriptors(CFloatImage &image, FeatureSet &features)
{
CFloatImage grayImage=ConvertToGray(image);
CFloatImage blurredImage;
Convolve(grayImage, blurredImage, ConvolveKernel_7x7);
CFloatImage postHomography = CFloatImage();
CFloatImage gaussianImage = GetImageFromMatrix((float *)gaussian5x5Float, 5, 5);
//first make the image invariant to changes in illumination by subtracting off the mean
int grayHeight = grayImage.Shape().height;
int grayWidth = grayImage.Shape().width;
// now make this rotation invariant
vector<Feature>::iterator featureIterator = features.begin();
while (featureIterator != features.end()) {
Feature &f = *featureIterator;
CTransform3x3 scaleTransform = CTransform3x3();
CTransform3x3 translationNegative;
CTransform3x3 translationPositive;
CTransform3x3 rotation;
double scaleFactor = 41/8;
scaleTransform[0][0] = scaleFactor;
scaleTransform[1][1] = scaleFactor;
translationNegative = translationNegative.Translation(f.x,f.y);
translationPositive = translationPositive.Translation(-4, -4);
rotation = rotation.Rotation(f.angleRadians * 180/ PI);
CTransform3x3 finalTransformation = translationNegative * rotation * scaleTransform * translationPositive;
//CFloatImage sample61x61Window =
//CFloatImage pixelWindow = GetXWindowAroundPixel(grayImage, f.x, f.y, 61);
WarpGlobal(blurredImage, postHomography, finalTransformation, eWarpInterpLinear, 1.0f);
//now we get the 41x41 box around the feature
for(int row=0; row< 8; row++)
{
for(int col=0;col< 8;col++)
{
f.data.push_back(postHomography.Pixel(col, row, 0));
}
}
/*
// now we do the subsampling first round to reduce to a 20x20
int imgSize = 41;
subsample(&f, imgSize, gaussianImage);
//second round of subsampling to get it to a 10x10
imgSize = 20;
subsample(&f, imgSize, gaussianImage);
imgSize = 10;
CFloatImage img = featureToImage(f, imgSize, imgSize);
CFloatImage blurredImg(img.Shape());
Convolve(img, blurredImg, gaussianImage);
featuresFromImage(&f,blurredImg,imgSize,imgSize);
int count = 0;
for(int y=0; y<imgSize; y++)
{
for(int x=0; x<imgSize; x++)
{
if(x == 3 || x == 7 || y == 3 || y == 7)
{
f.data.erase(f.data.begin() + count);
}
else
{
count++;
}
}
}
*/
normalizeIntensities(&f, 8, 8);
featureIterator++;
}
}
void InstantiateConvolutionOf(CImageOf<T> img)
{
CFloatImage kernel;
Convolve(img, img, kernel);
ConvolveSeparable(img, img, kernel, kernel, 1);
}
// Loop through the image to compute the harris corner values as described in class
// srcImage: grayscale of original image
// harrisImage: populate the harris values per pixel in this image
void computeHarrisValues(CFloatImage &srcImage, CFloatImage &harrisImage, CFloatImage &orientationImage)
{
int w = srcImage.Shape().width; // image width
int h = srcImage.Shape().height; // image height
// Create images to store x-derivative and y-derivative values
CFloatImage Ix(w,h,1);
CFloatImage Iy(w,h,1);
CFloatImage Ix_blur(w,h,1);
CFloatImage Iy_blur(w,h,1);
// Compute x-derivative values by convolving image with x sobel filter
Convolve(srcImage, Ix, ConvolveKernel_SobelX);
// Compute y-derivative values by convolving image with y sobel filter
Convolve(srcImage, Iy, ConvolveKernel_SobelY);
// Apply a 7x7 gaussian blur to the grayscale image
CFloatImage blurImage(w,h,1);
Convolve(srcImage, blurImage, ConvolveKernel_7x7);
// Compute x-derivative values by convolving blurred image with x sobel filter
Convolve(blurImage, Ix_blur, ConvolveKernel_SobelX);
// Compute y-derivative values by convolving blurred image with y sobel filter
Convolve(blurImage, Iy_blur, ConvolveKernel_SobelY);
// Declare additional variables
int newX, newY; // (x,y) coordinate for pixel in 5x5 sliding window
float dx, dy; // x-derivative, y-derivative values
double HMatrix[4]; // Harris matrix
double determinant; // determinant of Harris matrix
double trace; // trace of Harris matrix
int padType = 2; // select variable for what type of padding to use: , 0->zero, 1->edge, 2->reflect
// Loop through 'srcImage' and compute harris score for each pixel
for (int y = 0; y < h; y++) {
for (int x = 0; x < w; x++) {
// reset Harris matrix values to 0
memset(HMatrix, 0, sizeof(HMatrix));
// Loop through pixels in 5x5 window to calculate Harris matrix
for (int j = 0; j < 25; j++) {
find5x5Index(x,y,j,&newX,&newY);
if(srcImage.Shape().InBounds(newX, newY)) {
dx = Ix.Pixel(newX,newY,0);
dy = Iy.Pixel(newX,newY,0);
} else {
// Depending on value of padType, perform different types of border padding
switch (padType) {
case 1:
// 1 -> replicate border values
if (newX < 0) {
newX = 0;
} else if (newX >= w) {
newX = w-1;
}
if (newY < 0) {
newY = 0;
} else if (newY >= h) {
newY = h-1;
}
dx = Ix.Pixel(newX,newY,0);
dy = Iy.Pixel(newX,newY,0);
break;
case 2:
// 2 -> reflect border pixels
if (newX < 0) {
newX = -newX;
} else if (newX >= w) {
newX = w-(newX%w)-1;
}
if (newY < 0) {
newY = -newY;
} else if (newY >= h) {
newY = h-(newY%h)-1;
}
dx = Ix.Pixel(newX,newY,0);
dy = Iy.Pixel(newX,newY,0);
break;
default:
// 0 -> zero padding
dx = 0.0;
dy = 0.0;
break;
}
}
HMatrix[0] += dx*dx*gaussian5x5[j];
HMatrix[1] += dx*dy*gaussian5x5[j];
HMatrix[2] += dx*dy*gaussian5x5[j];
HMatrix[3] += dy*dy*gaussian5x5[j];
}
// Calculate determinant and trace of harris matrix
determinant = (HMatrix[0] * HMatrix[3]) - (HMatrix[1] * HMatrix[2]);
trace = HMatrix[0] + HMatrix[3];
// Compute corner strength function c(H) = determinant(H)/trace(H)
// and save result into harrisImage
if(trace == 0)
harrisImage.Pixel(x,y,0) = 0.0;
else
harrisImage.Pixel(x,y,0) = (determinant / trace);
// Compute orientation and save result in 'orientationImage'
dx = Ix_blur.Pixel(x,y,0);
dy = Iy_blur.Pixel(x,y,0);
if(dx == 0.0 && dy == 0.0)
orientationImage.Pixel(x,y,0) = 0.0;
else
orientationImage.Pixel(x,y,0) = atan2(dy, dx);
}
}
}
